API

spread(G::AbstractMatrix{Float64})

Calculate the transfer matrix for the adyacency matrix of the trilayered feature-source-target network.

Arguments

• G::AbstractMatrix{Float64}: Trilayered feature-source-target network adjacency matrix.

Extended help

Potential interactions between nodes in a graph can be identified by using resource diffusion processes in the feature-source-target network, namely aforementioned graph G. For each node nᵢ in the network, it has initial resources located in both its neighboring nodes and its features. Initially, each feature and each neighboring node of nᵢ equally spread their resources to neighboring nodes. Subsequently, each of those nodes equally spreads its resources to neighbor nodes. Thus, nᵢ will obtain final resources located in several neighboring nodes, suggesting that nᵢ may have potential interactions with these nodes.

References

1. Wu, et al (2016). SDTNBI: an integrated network and chemoinformatics tool for systematic prediction of drug–target interactions and drug repositioning. Briefings in Bioinformatics, bbw012. https://doi.org/10.1093/bib/bbw012
2. Vigil-Vásquez & Schüller (2022). De Novo Prediction of Drug Targets and Candidates by Chemical Similarity-Guided Network-Based Inference. International Journal of Molecular Sciences, 23(17), 9666. https://doi.org/10.3390/ijms23179666

cutoff(x::T, α::T, weighted::Bool=false) where {T<:AbstractFloat}

Transform x based in SimSpread's similarity cutoff function.

Arguments

• x::AbstractFloat : Value to transform
• α::AbstractFloat : Similarity cutoff
• weighted::Bool : Apply weighting function to outcome (default = false)

cutoff(M::AbstractVecOrMat{T}, α::T, weighted::Bool=false) where {T<:AbstractFloat}

Transform the vector or matrix X based in SimSpread's similarity cutoff function.

Arguments

• X::AbstractVecOrMat{AbstractFloat} : Matrix or Vector to transform
• α::AbstractFloat : Similarity cutoff
• weighted::Bool : Apply weighting function to outcome (default = false)

cutoff!(x::T, α::T, weighted::Bool=false) where {T<:AbstractFloat}

Transform, in place, x based in SimSpread's similarity cutoff function.

Arguments

• x::AbstractFloat : Value to transform
• α::AbstractFloat : Similarity cutoff
• weighted::Bool : Apply weighting function to outcome (default = false)

cutoff!(X::AbstractVecOrMat{T}, α::T, weighted::Bool=false) where {T<:AbstractFloat}

Transform, in place, the vector or matrix X based in SimSpread's similarity cutoff function.

Arguments

• X::AbstractVecOrMat{AbstractFloat} : Matrix or Vector to transform
• α::AbstractFloat : Similarity threshold
• weighted::Bool : Apply weighting function to outcome (default = false)

featurize(X::NamedArray, α::AbstractFloat, weighted::Bool=true)

Transform the feature matrix X into a SimSpread feature matrix.

Arguments

• X::NamedArray: Continuous feature matrix
• α::AbstractFloat: Featurization cutoff
• weighted::Bool : Apply weighting function to outcome (default = true)

References

1. Vigil-Vásquez & Schüller (2022). De Novo Prediction of Drug Targets and Candidates by Chemical Similarity-Guided Network-Based Inference. International Journal of Molecular Sciences, 23(17), 9666. https://doi.org/10.3390/ijms23179666

featurize!(X::NamedArray, α::AbstractFloat, weighted::Bool=true)

Transform, in place, the feature matrix X into a SimSpread feature matrix.

Arguments

• X::NamedArray : Continuous feature matrix
• α::AbstractFloat : Featurization cutoff
• weighted::Bool : Apply weighting function to outcome (default = true)

References

1. Vigil-Vásquez & Schüller (2022). De Novo Prediction of Drug Targets and Candidates by Chemical Similarity-Guided Network-Based Inference. International Journal of Molecular Sciences, 23(17), 9666. https://doi.org/10.3390/ijms23179666

construct(y::NamedMatrix, X::NamedMatrix, queries::AbstractVector)

Construct the query-feature-source-target network for de novo network-based inference prediction and return adjacency matrix.

Arguments

• y::NamedMatrix: Source-target bipartite network adjacency matrix
• X::NamedMatrix: Source-feature bipartite adjacency matrix
• queries::AbstractVector: Source nodes to use as query

Extended help

This implementation is intended for k-fold or leave-one-out cross-validation.

construct(ys::T, Xs::T) where {T<:Tuple{NamedMatrix,NamedMatrix}}

Construct the query-feature-source-target network for de novo network-based inference prediction and return adjacency matrix.

Arguments

• dts::Tuple{NamedMatrix,NamedMatrix} : Source-target bipartite graph adjacency matrices
• dfs::Tuple{NamedMatrix,NamedMatrix} : Source-feature bipartite graph adjacency matrices

Extended help

This implementations is intended for time-split cross-validation or manual construction of query network.

construct(ytrain::T, ytest::T, Xtrain::T, Xtest::T) where {T<:NamedMatrix}

Construct the query-feature-source-target network for de novo network-based inference prediction and return adjacency matrix.

Arguments

• ytrain::NamedMatrix : Training source-target bipartite graph adjacency matrix
• ytest::NamedMatrix : Test source-target bipartite graph adjacency matrix
• Xtrain::NamedMatrix : Training source-feature bipartite graph adjacency matrix
• Xtest::NamedMatrix : Test source-feature bipartite graph adjacency matrix

Extended help

This implementations is intended for time-split cross-validation or manual construction of query network.

construct(y::NamedMatrix, X::NamedMatrix)

Construct the feature-source-target network for network-based inference prediction and return adjacency matrix.

Arguments

• y::NamedMatrix : Source-target bipartite graph adjacency matrix
• X::NamedMatrix : Source-feature bipartite graph adjacency matrix

predict(I::Tuple{T,T}, ytest::T; GPU::Bool=false, returnweights::Bool=false) where {T<:NamedMatrix}

Predict interactions between query and target nodes using de novo network-based inference model proposed by Wu, et al (2016).

Arguments

• I::Tuple{NamedMatrix,NamedMatrix}: Feature-source-target trilayered adjacency matrices
• ytest::NamedMatrix: Query-target bipartite adjacency matrix
• GPU::Bool: Use GPU acceleration for calculation (default = false)
• returnweights::Bool: Return the weighting matrix employed for prediction

References

1. Wu, et al (2016). SDTNBI: an integrated network and chemoinformatics tool for systematic prediction of drug–target interactions and drug repositioning. Briefings in Bioinformatics, bbw012. https://doi.org/10.1093/bib/bbw012
2. Vigil-Vásquez & Schüller (2022). De Novo Prediction of Drug Targets and Candidates by Chemical Similarity-Guided Network-Based Inference. International Journal of Molecular Sciences, 23(17), 9666. https://doi.org/10.3390/ijms23179666

predict(A::T, ytrain::T; GPU::Bool=false, returnweights::Bool=false) where {T<:NamedMatrix}

Predict interactions between query and target nodes using de novo network-based inference model proposed by Wu, et al (2016).

Arguments

• A::NamedMatrix: Feature-source-target trilayered adjacency matrix
• ytrain::NamedMatrix: Source-target bipartite adjacency matrix
• GPU::Bool: Use GPU acceleration for calculation (default = false)
• returnweights::Bool: Return the weighting matrix employed for prediction

References

1. Wu, et al (2016). SDTNBI: an integrated network and chemoinformatics tool for systematic prediction of drug–target interactions and drug repositioning. Briefings in Bioinformatics, bbw012. https://doi.org/10.1093/bib/bbw012
2. Vigil-Vásquez & Schüller (2022). De Novo Prediction of Drug Targets and Candidates by Chemical Similarity-Guided Network-Based Inference. International Journal of Molecular Sciences, 23(17), 9666. https://doi.org/10.3390/ijms23179666

save(filepath::String, yhat::NamedMatrix, y::NamedMatrix; delimiter::Char='	')

Store predictions as a table in the given file path.

Arguments

• filepath::String: Output file path
• yhat::NamedArray: Predicted source-target bipartite adjacency matrix
• y::NamedArray: Ground-truth source-target bipartite adjacency matrix
• delimiter::Char: Delimiter used to write table (default = '\t')

Extended help

Table format is:

fold, source, target, score, label

save(filepath::String, fidx::Int64, yhat::NamedMatrix, y::NamedMatrix; delimiter::Char='	')

Store cross-valudation predictions as a table in the given file path.

Arguments

• filepath::String: Output file path
• fidx::Int64: Numeric fold ID
• yhat::NamedArray: Predicted source-target bipartite adjacency matrix
• y::NamedArray: Ground-truth source-target bipartite adjacency matrix
• delimiter::Char: Delimiter used to write table (default = '\t')

Extended help

Table format is:

fold, source, target, score, label

Base.split(y::NamedArray, k::Int64; seed::Int64=1)

Split source nodes in y into k groups for cross-validation.

Arguments

• y::AbstractMatrix: Drug-Target rectangular adjacency matrix.
• k::Int64: Number of groups to use in data splitting.
• seed::Int64: Seed used for data splitting.

clean!(yhat::NamedArray, A::NamedArray, y::NamedArray)

Flag, in place, erroneous prediction from cross-validation splitting.

Arguments

• yhat::NamedArray: Predicted source-target bipartite adjacency matrix
• A::NamedArray: Initial resource source-target resources adjacency matrix
• y::NamedArray: Ground-truth source-target bipartite adjacency matrix

AuPRC(y::AbstractVector{Bool}, yhat::AbstractVector)

Area under the Precision-Recall curve using the trapezoidal rule.

Arguments

• y::AbstractArray: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractArray: Prediction scores.

AuROC(y::AbstractVector{Bool}, yhat::AbstractVector)

Area under the Receiver Operator Characteristic curve using the trapezoidal rule.

Arguments

• y::AbstractArray: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractArray: Prediction scores.

recallatL(y, yhat, L)

Get recall@L as proposed by Wu, et al (2017).

Arguments

• y::AbstractVector: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractVector: Prediction score.
• L::Integer: Length to consider to calculate metrics (default = 20).

recallatL(y, yhat, L)

Get mean recall@L per group as proposed by Wu, et al (2017).

Arguments

• y::AbstractVector: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractVector: Prediction score.
• ̂grouping::AbstractVector: Group labels.
• L::Integer: Length to consider to calculate metrics (default = 20).

precisionatL(y, yhat, L::Integer=20)

Get precision@L as proposed by Wu, et al (2017).

Arguments

• y::AbstractVector: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractVector: Prediction score.
• L::Integer: Length to consider to calculate metrics (default = 20).

precisionatL(y, yhat, grouping, L)

Get mean precision@L per group as proposed by Wu, et al (2017).

Arguments

• y::AbstractVector: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractVector: Prediction score.
• grouping::AbstractVector: Group labels.
• L::Integer: Length to consider to calculate metrics (default = 20).

BEDROC(y::AbstractVector{Bool}, yhat::AbstractVector; rev::Bool=true, α::AbstractFloat=20.0)

The Boltzmann Enhanced Descrimination of the Receiver Operator Characteristic (BEDROC) score is a modification of the Receiver Operator Characteristic (ROC) score that allows for a factor of early recognition.

Score takes a value in interval [0, 1] indicating degree to which the predictive model employed detects (early) the positive class.

Arguments

• y::AbstractArray: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractArray: Prediction scores.
• rev::Bool: True if high values of $yhat$ correlates to positive class (default = true).
• α::AbstractFloat: Early recognition parameter (default = 20.0).

References

1. Truchon, J.-F., & Bayly, C. I. (2007). Evaluating Virtual Screening Methods:  Good and

Bad Metrics for the “Early Recognition” Problem. Journal of Chemical Information and Modeling, 47(2), 488–508. https://doi.org/10.1021/ci600426e

f1score(tn::T, fp::T, fn::T, tp::T) where {T<:Integer}

The harmonic mean between precision and recall

Arguments

• tn::Integer True negatives
• fp::Integer False postives
• fn::Integer False negatives
• tp::Integer True positives

mcc(a::T, b::T, ϵ::AbstractFloat = 0.0001) where {T<:Integer}

Matthews correlation coefficient using calculus approximation for when FN+TN, FP+TN, TP+FN or TP+FP equals zero.

Arguments

• a::Integer = Value of position a in confusion matrix
• b::Integer = Value of position b in confusion matrix
• ϵ::AbstractFloat = Approximation coefficient (default = floatmin(Float64))

Extended help

The confusion matrix in a binary prediction is comprised of 4 distinct positions:

                    | Predicted positive     Predicted negative
    ----------------+--------------------------------------------
    Actual positive |  True positives (TP)   False negatives (FN)
    Actual negative | False positives (FP)    True negatives (TN)

In the case a row or column of the confusion matrix equals zero, MCC is undefined. Therefore, to correctly use MCC with this approximation, arguments a and b are defined as follows:

• If "Predictive positive" column is zero, a is TN and b is FN
• If "Predictive negative" column is zero, a is TP and b is FP
• If "Actual positive" row is zero, a is TN and b is FP
• If "Actual negative" row is zero, a is TP and b is FN

Reference

1.Chicco, D., Jurman, G. The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation. BMC Genomics 21, 6 (2020).

mcc(tn::T, fp::T, fn::T, tp::T) where {T<:Integer}

Matthews correlation coefficient, a special case of the phi coeficient Performance metric used for overcoming the class imbalance issues

Arguments

• tn::Integer True negatives
• fp::Integer False postives
• fn::Integer False negatives
• tp::Integer True positives

Reference

1.Chicco, D., Jurman, G. The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation. BMC Genomics 21, 6 (2020).

accuracy(tn::T, fp::T, fn::T, tp::T) where {T<:Integer}

The number of all correct predictions divided by the total predicitions

Arguments

• tn::Integer True negatives
• fp::Integer False postives
• fn::Integer False negatives
• tp::Integer True positives

balancedaccuracy(tn::T, fp::T, fn::T, tp::T) where {T<:Integer}

The arithmetic mean of sensitivity and specificity, its use case is when dealing with imbalanced data

Arguments

• tn::Integer True negatives
• fp::Integer False postives
• fn::Integer False negatives
• tp::Integer True positives

recall(tn::T, fp::T, fn::T, tp::T) where {T<:Integer}

The fraction of positive samples correctly predicted as postive

Arguments

• tn::Integer True negatives
• fp::Integer False postives
• fn::Integer False negatives
• tp::Integer True positives

precision(tn::T, fp::T, fn::T, tp::T) where {T<:Integer}

The fraction of positive predictions that are correct

Arguments

• tn::Integer True negatives
• fp::Integer False postives
• fn::Integer False negatives
• tp::Integer True positives

meanperformance(confusion::AbstractVector{ROCNums{Int64}}, metric::Function)

Get mean performance of a given metric over a set of confusion matrices.

Arguments

• confusion::AbstractVector{ROCNums{Int64}}: Confusion matrices
• ̂metric::Function: Performance metric function to use in evaluation.

meanperformance(y::AbstractVector{Bool}, yhat::AbstractVector{Float64}, metric::Function)

Get mean performance of a given metric over a pair of label-prediction vectors.

Arguments

• y::AbstractVector: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractVector: Prediction score.
• ̂metric::Function: Performance metric function to use in evaluation.

meanstdperformance(confusion::AbstractVector{ROCNums{Real}}, metric::Function)

Get mean and standard deviation performance of a given metric over a set of confusion matrices.

Arguments

• confusion::AbstractVector{ROCNums{Real}}: Confusion matrix object from MLBase
• ̂metric::Function: Performance metric function to use in evaluation.

meanstdperformance(y::AbstractVector{Bool}, yhat::AbstractVector{Float64}, metric::Function)

Get mean and standard deviation performance of a given metric over a pair of label-prediction vectors.

Arguments

• y::AbstractVector: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractVector: Prediction scores.
• ̂metric::Function: Performance metric function to use in evaluation.

maxperformance(confusion::AbstractVector{ROCNums{Real}}, metric::Function)

Get maximum performance of a given metric over a set of confusion matrices.

Arguments

• confusion::AbstractVector{ROCNums{Real}}: Confusion matrices.
• ̂metric::Function: Performance metric function to use in evaluation.

maxperformance(y::AbstractVector{Bool}, yhat::AbstractVector{Float64}, metric::Function)

Get maximum performance of a given metric over a pair of label-prediction vectors.

Arguments

• y::AbstractVector{Bool}: Binary class labels. 1 for positive class, 0 otherwise.
• ̂yhat::AbstractVector{Float64}: Prediction score.
• ̂metric::Function: Performance metric function to use in evaluation.

validity_ratio(yhat::AbstractVector)

Ratio of valid predictions (score > 0) and all predictions. Allows to check if predictive performance is given for all predictions or a subset of the predictions.

Arguments

• yhat::AbstractVector : Prediction scores.

Extended help

A limitation of SimSpread is that it is impossible to generate predictions for query nodes whose similarity to every source of the first network layer is below the similarity threshold α. In this case, the length of the feature vector is zero, resource spreading is not possible, and the predicted value is zero for all targets, therefore we need guardrails to correctly assess performance as a function of the threshold α.

This characteristic of SimSpread can be seen as an intrinsic notion of its application domain. No target predictions are generated for query nodes outside SimSpread’s application domain instead of returning likely meaningless targets.

References

1. Vigil-Vásquez & Schüller (2022). De Novo Prediction of Drug Targets and Candidates by Chemical Similarity-Guided Network-Based Inference. International Journal of Molecular Sciences, 23(17), 9666. https://doi.org/10.3390/ijms23179666

read_namedmatrix(filepath::String, valuetype::Type = FLoat64filepath::String, valuetype::Type)

Load a matrix with named indices as a NamedArray.

Arguments

• filepath::String : File path of matrix to load
• delimiter::Char : Delimiter character between values in matrix (default = ' ')
• valuetype::Type : Type of values contained in matrix (default = Float64)

k(G::AbstractMatrix)

Get node degrees from adjacency matrix

Arguments

• M::AbstractMatrix : Matrix to parse

getyamanishi(db)

Get a tuple of matrices corresponding to the drug-target adjacency matrix and drug-drug similarity matrix for a given Yamanishi (2008) dataset.

Arguments

• db: Dataset ID (any of the following: "nr", "ic", "gpcr" or "e")

Example

julia> dt, dd = getyamanishi("nr");

julia> dt[1:5, 1:5]
5×5 Named Matrix{Float64}
 A ╲ B │  hsa190  hsa2099  hsa2100  hsa2101  hsa2103
───────┼────────────────────────────────────────────
D00040 │     0.0      0.0      0.0      0.0      0.0
D00066 │     0.0      1.0      0.0      0.0      0.0
D00067 │     0.0      1.0      0.0      0.0      0.0
D00075 │     0.0      0.0      0.0      0.0      0.0
D00088 │     0.0      0.0      0.0      0.0      0.0

julia> dd[1:5, 1:5]
5×5 Named Matrix{Float64}
 A ╲ B │   D00040    D00066    D00067    D00075    D00088
───────┼─────────────────────────────────────────────────
D00040 │      1.0  0.545455  0.297297   0.53125  0.459459
D00066 │ 0.545455       1.0  0.387097  0.833333  0.689655
D00067 │ 0.297297  0.387097       1.0  0.464286  0.352941
D00075 │  0.53125  0.833333  0.464286       1.0  0.678571
D00088 │ 0.459459  0.689655  0.352941  0.678571       1.0

Extended help

The provided Yamanishi (2008) [1] datasets (ID) are:

• 'Nuclear Receptor' (nr)
• 'Ion Channels' (ic)
• 'GPCR' (gpcr)
• 'Enzyme' (e)

This function returns 2 distinct adjacency matrices:

• Binary drug-target interaction matrix, obtained from biological annotations
• Continuous drug-drug similarity matrix, obtained from SIMCOMP

References

1. Yamanishi, Y., Araki, M., Gutteridge, A., Honda, W., & Kanehisa, M. (2008). Prediction of drug–target interaction networks from the integration of chemical and genomic spaces. Bioinformatics, 24(13), i232–i240. https://doi.org/10.1093/bioinformatics/btn162